FDTransUnet: An aluminum surface defect segmentation model based on feature differentiation.

Area of Science:

• Computer Vision
• Machine Learning
• Materials Science

Background:

• Industrial defect segmentation faces challenges with limited data, low recognition accuracy, and poor segmentation precision.
• Existing methods struggle with insufficient sample sizes, leading to overfitting and reduced model performance.
• Accurate defect detection is crucial for quality control in manufacturing processes.

Purpose of the Study:

• To propose FDTransUnet, a surface defect segmentation model for aluminum utilizing feature differentiation.
• To address data scarcity, improve recognition and segmentation accuracy, and enhance model generalization.
• To develop a robust solution for industrial inspection scenarios.

Main Methods:

• Employed a feature differentiation data augmentation strategy to expand limited defective samples and mitigate overfitting.
• Integrated Transformer architecture into U-net, combining global self-attention with hierarchical structure for effective information extraction.
• Constructed a composite loss function to handle foreground-background class imbalance and boost segmentation accuracy.

Main Results:

• FDTransUnet achieved 94.5% Mean Pixel Accuracy (MPA) and 89.7% Dice coefficient on an aluminum surface defect dataset.
• Generalization experiments on a steel surface defect dataset demonstrated FDTransUnet's strong performance compared to mainstream models.
• The model exhibited good generalization performance and robustness across different industrial inspection scenarios.

Conclusions:

• FDTransUnet effectively overcomes data limitations and improves surface defect segmentation accuracy.
• The hybrid U-net Transformer architecture and composite loss function contribute to enhanced performance.
• The proposed model shows significant potential for real-world industrial inspection applications.